but i wanted to presented to a continuum of what nickel are presented them
are you
about
we created a an inquisition game
that abstracted completely the natural language out from the game so basically we consider that
the negotiation
it's
it's finding an agreement between two people
that's not my
it's fa an agreement between two people and
and so both there are a number of options possible agreement between them
both people have different preferences overall this set of possible agreements
and they have to exchange in order to find the best possible way
and what we found out
so you know very simply for the you know very simplified way of
of communicating the different options is that supposed to simple and when we put two
humans
trying to find an agreement together
maybe they didn't the where some of them where more easygoing
but in the n the best strategy for the system was quite the same
so the continue the
the well what we did what we're trying to do now is to of more
complex options that are
that's not the combination of several features so for instance issue of trying to set
an appointment you have to specify the date you have to specify the power
or if you want to exchange fruits like in the bargaining task
well you have to define how many apple i mean the oranges you want and
stuff like that so these are different features and you can have a much complex
a set of actions
that a i want the apples
or a
we can meet in the at any time on a on thursday
stuff like this and it
the number of action then explodes and it would be much more in an interesting
to work on this
but i actually need
i other than that huh show and the unit can try to
i mean they can start like to tease to show the example
i know it was the presentation and the k
so we start the
so
i think that everybody here agrees that major challenge for the automatic analysis them negotiate
negotiation dialogs
is that a like modeling be disagreement space which is shaped it different participants in
you process of arguing
and actually at least it's of the art there like studied in argumentation mining that
focuses on due to medic identification of like james and primacy is and all sounded
types of relation linking them like that supports
agreement or disagreement
our have are the current methodologies do not find to what aspects easy to the
treatment as scope over
and we believe that this is actually important in order to predict registration strategies and
also to understand like specific controversy in different contexts
so therefore our research question is like how can we model these scope of disagreements
you know comment that the context
and on these grounds we proposed a to a level ontology an upper lever and
the lower level ontology
in order to model these agreement space so let's like take as example
a discussion around like taken from change maybe you the a subgradient
and so as you can see details all of the original post use like diversity
is not about race
and uncertainty common starts to be at i versus societies the society which have people
from different backgrounds and cultures how does raise scamming to play
and actually these two sentences are called out or this is how we call nh
comments in our like upper level ontology by one random participant that is called like
d in the skin
that actually challenges like the assertion underlying the rhetorical question we do not a rhetorical
question so you're joking right and you can do you know as we descended like
talk to some black falls
so it's clears the according to our ontology did these first to send its use
in like orange in the original forest
are a target
and the comment is like a whole lot
and so it is clear that the relation is that these agreements relation but actually
you can i think an e like weekly understand what is challenge is not really
like the each row of the statement it is the fact that it's the each
roles of the person with like expressing statement
so basically what is challenge is not be proposition but speech act so one pretty
easy to use conditional like making an argument is that of like having your right
to do so
and so actually the user is claiming that like the speaker is biased
then you original poster goes on and you like you provides an example take these
example a white child we immigrated from change are yet this depressed persecution and the
black and make an child even next door to each other
the african american child was had another way someone childhood is accepted into college for
this take a bigger city one channel from chain child that is rejected
so in this case there is an adder user id that like calls out this
time like the challenge is you know corrine somebody rejection event
and it is like well it's hard to say which key it would have a
better shot at getting into the same competitive school and then you can tune you
like expressing like a these agreement again but towards like the last sentence of the
original what was so that prosody should be about the result of experience in background
not skin
so in this case what is challenge is not the events but actually it's really
like the you truth of deeper position
another type of these agreement is they one expressed by be all actually like in
is
commenting on d meaning of the verb novelty word by diversity
which is also part of the last a statement which actually in like different linguistic
theories is considered like an entity soft first order and d
so there are also other types of relation you can also have like agreement relations
for example like the original poster answers to user c and this is like no
i agree
that one what it was like trying to say or a to relations type that
are hard to classify
for example like when the what we jump was to a of words i delta
so you e actually that they like
you have or stating
so
yes i just want to finish we award punch lines our pension is that these
ontologies readable to leverage outer existing semantic and pragmatic layers of annotation
and then reach the information the information that they provide and for example because relation
between like these agreement relations and you types of targets that they select
promises to improve the detection of like about those which are like a type of
disagreement
and so then we then we like these are like a more points of discussion
that basically muir those that have already been addressed by dependencies
thank you
so i'm happy to announce it just happens like two weeks ago via released out
the corpus and you corpus met the log multi-issue bargaining corpus in ldc catalog so
it's
i presented for all for our group in here in several on and that colleagues
in grounding and cool build the cognitive model for the corpus for the corpus collection
and for the future system and i will be present the system to model
so what the corpus is about as a scenario is multi-issue bargaining so it's not
just negotiation to buy banana or oranges is i issues based preferences involve therefore issuers
it's integrated negotiation as win been situation
it's featuring actually negotiation the value in my it's a complex negotiation strategic negotiation the
domain it was real scenario took
in a at the necessity of buttons past the anti smoking legislation also this each
year of new york could force the debate on the on this one and it
was not very efficient so they me to come immediate many by just need to
come together and everything negotiate
why is not walking a half adjustments so basically the corpus is collected to be
is that a negotiation train there so dct council needs to train a to be
trained to negotiate beginnings different body sit giddings business represent the thief
against police officers against house insurance et cetera et cetera so the preferences brag even
for them raise up references in the in a sense that the right couple of
scenarios designs and because
it was not real politicians involve by the out you again parliamentarians so that where
got a preferences and they need to defend their positions no time constraints and they
were instructed to a weight negotiate a negative agreement i will explain later tomorrow so
the basically
cannot be all they we're at college not to accept
this preferred
this preferred options
so the we will release for parts of this corpus now we release human dialogues
for to have a was eight subjects
two thousand turns ten thousand tokens bill the release they next part more which is
more argument that eve we had they need to defend their position we will release
debate corpus on the same topic and we will release the evaluation corpus which is
larger human
machine dialogue or human machine dialogues
corpus asr recordings and transcriptions and you can use these to retrain the speech recognition
is obviously too small but you can use it for adaptations of the speaker diarisation
is done manually correctly everything a high quality
into format in many format's also dear and also transcriptions in t i is include
eats
down we have dialogue annotations or semantic and pragmatic annotations
about nine cells and to vent that this out annotated this types six dives up
before the dialogue acts discussed structuring acts rhetorical relations according to the newest eyes are
stand that
that's it and this is you your l where you can download the
corpus if you have membership